AI News, Oct 5: OpenAI Moves Ads Into Image Generation
OpenAI is putting its first visual ad format inside ChatGPT’s image generator, where 1.2 billion weekly users make things. Hacker News spent the same 24 hours on two tools: one for running a 125-billion-parameter model on a gaming GPU, and one for deleting Apple Intelligence.
The Big Story: The Ad Moved Inside the Thing You’re Making
OpenAI will show visual ads in ChatGPT while you generate images, and says it will “begin testing the new ad format later this month in the U.S. with an initial group of advertisers.” Ads “will be clearly labeled, and remain separate from the image being created,” the company says, and it reiterated that ads do not influence ChatGPT’s answers. It also says it will avoid placements in emotionally sensitive conversations. ChatGPT reaches “1.2 billion people every week.”
This is not OpenAI introducing ads. ChatGPT has carried them since February on the free and Go tiers in the US, as labeled units in the conversation, shopping carousels and sponsored search results, and a self-serve ads manager opened in May. What changed is where the ad sits. Every previous format put the sponsored unit next to the answer, preserving a seam between what you asked for and what someone paid to show you. Image generation has no such seam. You watch a canvas waiting for a picture to resolve, and now a paid visual appears in the same frame.
The measurement stack says more than the announcement does. OpenAI named Hightouch, Tealium and LiveRamp for conversion, AppsFlyer, Adjust, Branch, Triple Whale and Kochava for attribution, and DoubleVerify and Integral Ad Science for brand safety. That is not an experiment’s worth of plumbing. It is the apparatus of a real ad network, assembled ahead of the inventory and pointed at the one surface where a user is demonstrably in a commercial mood.
Hacker News spent the weekend running AI locally and ripping it out of macOS
The two biggest stories on Hacker News in 24 hours, by a wide margin, were a tool for running a near-frontier model yourself and a tool for removing a vendor’s AI from your laptop. Between them they took nearly 1,500 points and more than 800 comments. Every other AI story in the window stayed under 100 points.
Strata hit 844 points and 371 comments for serving Qwen3.8-Flash-Next, a 125B mixture-of-experts model, on a single consumer card with 12GB of VRAM via expert offload. On an RTX 5070 the README measures 94 tokens a second of generation and 2,650 of prompt processing at Q2_0, the developer’s own numbers on their own gaming PC with no third-party reproduction. The submitted headline’s “RTX 4090 at 100T/s” appears nowhere in the repo, which has no 4090 row at all. What matters architecturally is that Strata exposes OpenAI- and Anthropic-compatible endpoints on localhost, so it drops straight into coding agents already pointed at a cloud API. The sharpest critique in the thread: nobody publishes benchmarks for the quantized weights, so the effective intelligence is unknown.
RemoveMacAI took 645 points and 439 comments for the opposite job. It is a reversible CLI that turns off Siri, Writing Tools, Genmoji, Image Playground, the ChatGPT extension, mail and message summaries and Xcode predictive completion, then deletes the on-device models and reclaims the disk. Half the thread argued about the curl | bash installer. The more interesting half tied the tool to macOS 27’s new Full Disk Access prompts breaking agentic workflows.
A Stratechery essay on Apple and agents drew 86 points the same morning and split the same way: reasonable security, or platform control in a security costume. Three of the window’s loudest threads are about who decides what AI runs on your machine. The lead story is about an ad appearing in a canvas you thought was yours.
Today’s Top Stories
Qualcomm will pay Huawei for 5G and AI patents, reversing 25 years
Huawei and Qualcomm have struck a cross-licensing agreement covering “5G and AI technologies” under which Qualcomm will pay to use Huawei’s patents, reversing a flow that has run the other way since Huawei first licensed Qualcomm’s patents in 2001. Nikkei calls it “a breakthrough for the Chinese company’s intellectual property portfolio.” The deal will trigger a regulatory review in the US and needs approval to proceed. No royalty figure or term was disclosed, and Nikkei is so far the only outlet carrying it. A US chipmaker paying a sanctioned Chinese firm for AI patents is a licensing story that reads as a capability story.
Korea is buying 10,000 GPUs to beat China’s open weights
Korea’s Ministry of Science and ICT will put 4.7 trillion won, about $3.48 billion, of equity investment into a frontier AI project starting next year, explicitly aimed at a model “capable of competing with leading Chinese open-source models.” The split is 3.9 trillion won for 10,000 Nvidia Vera Rubin GPUs and 800 billion won for training data, with participating companies required to match the government’s money privately, possibly through a special purpose company. Equity rather than grants, so returns flow back to the public.
It also ends an experiment. The existing national foundation-model programme, launched in 2025 as survival-style elimination and now down to three teams with two finalists due by year end, shifts toward industry-specific models. “The competition created a stigma for companies that were eliminated and stirred public controversy,” Second Vice Minister Ryu Je-myung said. Asked whether to keep handing out GPUs and letting firms race, he said the companies themselves wanted the resources pooled.
AI-generated child abuse imagery rose 40% in half the time
The Internet Watch Foundation assessed 6,310 AI images meeting the legal definition of child sexual abuse in the first half of 2026, 40% higher than last year’s full-year total of more than 4,500. The majority were of girls aged seven to 13. Its Report Remove service has taken 420 reports from children about faked or manipulated images of themselves, already past 397 for all of 2025.
UK AI minister Kanishka Narayan has said “nothing is off the table” on fresh regulation, with the test being “Will it make the British people safer?” The IWF’s head of policy, Hannah Swirsky, backed calls for binding legislation: “It is incumbent on tech companies to build tools which cannot be abused this way.” Britain has already made it a crime to adapt or distribute a model for this purpose, with up to five years for developing one. Half-year numbers beating a full prior year is the shape of a problem paced by generation capacity, not demand.
Anthropic read a user’s diary and called the police
A Florida woman faces a second-degree felony charge after Anthropic reported what she wrote to Claude. Carli Michelle Heller of Bonita Springs, who used Claude “like a ‘diary’,” wrote on September 26 that she would shoot up the Sheriff’s office. Claude’s safety systems flagged the entry, a human reviewer judged it a credible threat, and deputies detained her without incident. She is charged under Florida Statute 836.10, which covers sending, posting or transmitting “a written or electronic record” threatening violence. Anthropic says it may share user information in limited emergencies if disclosure is necessary to prevent death or serious physical injury.
The Hacker News thread went straight to the statute, arguing prosecutors are stretching a law about records that may be viewed by another person to cover a private conversation with a model. The strongest counter granted Anthropic a damned-either-way position, given the backlash when a rival failed to escalate a similar case. A chatbot is a Big Tech product, not a confidant, and unlike a notebook it is screenable at planetary scale.
AI took 64% of every venture dollar on earth last quarter
Global venture funding hit $159 billion in Q3 2026 across close to 6,000 startups, and AI companies took $102 billion of it, or 64%. The quarter was down 25% from Q2’s $212 billion but up 53% against Q3 2025’s $104 billion. Concentration is the striking figure: an all-time record 27 companies raised billion-dollar rounds, against a previous record of 16 in Q2 and 14 in Q1, and roughly a third of all global venture capital in the quarter went to those 27. Through three quarters, 2026 has absorbed $679 billion, the highest first-three-quarters total on record.
Quick Hits
- Model releases: Of more than fifty lab and vendor namespaces checked on Hugging Face, exactly two uploaded anything in the window. Of the 3,098 models created there overall, everything with download traction was a quantization mirror.
- Meeting transcripts: The one real release was Google’s DiarizationLM-Gemma-4-E4B-v1, a 4B speaker-diarization fine-tune that cuts word diarization error on Fisher from 5.32% to 2.99% and beats its own Llama-3 8B predecessor at half the parameters. The figures are the authors’ own, and the card says it is “not an officially supported Google product.”
- Meeting notes: Carly connects to thousands of apps, pulls in notes from all of your notetakers, and can start a workflow when something happens in one of them.
- Lab politics: Sam Altman refused the trade of capability for safety in a Politico interview, saying “I wouldn’t take a trade of saying, ‘we’ll make sure there’s no major hacks, there’s no misuse of this technology, there’s zero scams, there’s zero all the other bad things that will happen,’” and putting “a lot of daylight” between himself and Dario Amodei over whether one San Francisco lab should hold the technology, which he called “a completely unacceptable trade-off.”
- Founder as subject: In a two-part Vanity Fair interview Altman called Elon Musk a bully, said “once in a while you got to punch a bully back in the nose,” and said that if mission and share price ever conflict, “we are not going to pay attention to the share price.”
- Open weights: Axios reports Nvidia-backed Reflection is preparing its first open-weight model, expected to lag the top US systems but compete with the best Chinese open weights, with other Western releases following this month. No name, parameter count or benchmark yet.
- Robotics: Panasonic’s chief AI officer told Nikkei it will begin making humanoid robots by 2029 for plants and warehouses, leaning on its battery business for the power system. Kawasaki named a 2030 target last week, so this is a race with no products in it.
- Robot safety: Safeworld left stealth with more than $12 million led by Shine Capital and a16z Speedrun, running robot control systems through simulated blind-corner collisions and tripping hazards. “How do you underwrite the risk of a probabilistic system?” asks co-founder Ding Zhao, who directs Carnegie Mellon’s Safe AI Lab.
- Agentic security: Paris-based Fleuret AI raised €4 million pre-seed led by RAISE Ventures for two agents, Emile and Champollion, that turn penetration testing into a continuous process and run on European infrastructure.
- Legal research: Copenhagen’s Pandektes raised €13.5 million led by Alstin Capital, with 500-plus customers across Denmark, Germany and Switzerland and revenue up sixfold over the past year.
- Defence autonomy: Buntar Aerospace raised $16 million led by Freedom Fund to scale its Buntar-3 reconnaissance drone for the Armed Forces of Ukraine, taking it to $30 million raised.
- Robotics unicorn: Munich’s RobCo crossed a $1 billion valuation on a $40 million-plus employee share sale, roughly double its January Series C mark, with Cherry Ventures and European Tech Collective buying in alongside Sequoia and Lightspeed.
- Data centers: A reporter copy-pasted text Google had marked trade secret in a Nebraska filing, revealing that its Lincoln facility reported 52.65 megawatts at peak demand and 13.299 megagallons of water over the past year, less than half what the City of Lincoln itself reported using on a single day in September. The story drew the window’s highest comment count on Hacker News, where top replies were notably contrarian about the datacenter-water narrative.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
